PSO based data clustering with a different perception

Abstract

Generally, the Particle Swarm Optimization (PSO) algorithm has two memory dimensions: cognitive and social. In this study, a new dimension called family memory has been introduced to enhance PSO-based clustering's performance wherein not only the data were clustered but also the particles. Earlier studies on PSO-based clustering had used uniform distribution to create the particles' coordinates. In contrast, a multi-normal distribution-based swarm initialization was proposed in this study wherein K-means clustering outcomes were used as an input, which ensured the relationship within a particle's coordinates. The performance improvement through each modification to the Conventional PSO (CPSO) algorithm was demonstrated, and the statistical tests proved the same. Based on the two proposed changes, a modified PSO (MPSO) algorithm was developed. To extend the validation, both the proposed changes were incorporated into four different PSO variants, and the performance improvements were tested. Further, the MPSO algorithm was compared with the four different clustering algorithms (three algorithms are evolutionary type, and one is a non-evolutionary type) present in the recent literature. Statistical tests for the same proved the MPSO’s significance. As an application, the MPSO algorithm was applied to Multi-criteria Inventory classification (MCIC) data. Its classification was compared with the nineteen recent inventory models’ classification using the total relevant cost, fill rate, total safety stock cost, and the validity indices. Besides, a new index has been developed to identify the most similar models to MPSO.